Background masking with SAM 3¶
What to look for: everything outside the mask is black in the model input, so a correspondence can only land on the animal.
Before feature extraction every image is background-masked, so keypoints land on the animal rather than on the vegetation or ground that stationary camera traps repeat. Where a dataset ships masks they are used; otherwise the animal is segmented with text-prompted SAM 3 and all detected instances are merged into one mask. This demo runs SAM 3 with the prompt "animal" on two groups of images:
- Synthetic renders: the twenty renders of the synthetic matching demo. The synthetic subset ships segmentation masks, so each SAM 3 mask is compared with them.
- Camera-trap and field photographs: the query and top-1 photo of every dataset in the paper, the same pairs as the before/after demo. Only CzechLynx ships segmentation masks, so the comparison appears for its two photos; for the other datasets the masks used in the paper were computed by us and the demo shows the SAM 3 mask alone.
- Divider: raw image on the left, masked model input on the right. Everything outside the mask is set to black; nothing is cropped or recentred.
- Outlines: red is the SAM 3 mask; blue, where available, the segmentation mask shipped with the dataset.
- Readouts: the text prompt actually used, SAM 3's detection confidence, how many instances were merged, the foreground share of the frame, and, where a dataset mask exists, the intersection over union with it.
- Where the masks disagree: on snow renders the dataset mask includes the animal's cast shadow and SAM 3 does not. Sea stars are not found by the prompt "animal" at all; the pipeline's fallback prompt "sea star" then segments them, and the readout shows it.
Synthetic lynx renders from the CzechLynx synthetic subset (Picek et al.), Zenodo record 17592004, CC BY 4.0. Photographs from CzechLynx, WildlifeReID-10k and SalamanderID2025 (AnimalCLEF 2025), the latter shown with the dataset team's permission.